softmax, categorical, multinomial with K-1 parameters for identifiability

Add new versions of all these functions for the (K-1)-vector parameterization of softmax.

softmax_dofm1(vector x)  =def=  softmax(c(x,0) )

where c(x,0) is the R function to append 0 to the end of x.

The advantage is that the degrees of freedom matches that of the simplex target, so the model's identified without priors.

The disadvantage is that the priors have to change relative to the K-value parameterization and are no longer symmetric.

• Any suggestions for a better name than dofm? I suppose we could say identified.
• Any preference for c(x,0) or c(0,x), i.e., should the 0 value be the first or the last element?